Trajectory-Based Data Forwarding for Light-Trafﬁc … Data Forwarding for Light-Trafﬁc Vehicular Ad-Hoc Networks Jaehoon Jeong, Student Member, IEEE, Shuo Guo, Student Member,

Trajectory-Based Data Forwarding forLight-Traffic Vehicular Ad-Hoc Networks

Jaehoon Jeong, Student Member, IEEE, Shuo Guo, Student Member, IEEE, Yu Gu, Student Member, IEEE,Tian He, Member, IEEE , and David H.C. Du, Fellow, IEEE

✦

Abstract—This paper proposes a Trajectory-Based Data Forwarding (TBD)scheme, tailored for the data forwarding for road-side reports in light-trafficvehicular ad-hoc networks. State-of-the-art schemes have demonstratedthe effectiveness of their data forwarding strategies by exploiting knownvehicular traffic statistics (e.g., densities and speeds). These results areencouraging, however, further improvements can be made by taking advan-tage of the growing popularity of GPS-based navigation systems. This paperpresents the first attempt to effectively utilize vehicles’ trajectory informationin a privacy-preserving manner. In our design, such trajectory information iscombined with the vehicular traffic statistics for a better performance. In adistributed way, each individual vehicle computes its end-to-end expecteddelivery delay to the Internet access points based on its position on itsvehicle trajectory and exchanges this delay with neighboring vehicles todetermine the best next-hop vehicle. For the accurate end-to-end delaycomputation, this paper also proposes a link delay model to estimate thepacket forwarding delay on a road segment. Through theoretical analysisand extensive simulation, it is shown that our link delay model provides theaccurate link delay estimation and our forwarding design outperforms theexisting scheme in terms of both the data delivery delay and packet deliveryratio.

Index Terms—Vehicular Network, Road Network, Data Forwarding, Trajec-tory, Link Delay, Delivery Delay.

1 INTRODUCTION

With the standardization of Dedicated Short Range Commu-nication (DSRC) by IEEE [1], Vehicular Ad Hoc Networks(VANETs) have recently reemerged as one of promisingresearch areas for safety and connectivity in road networks.Currently, most research and development fall into one oftwo categories: (i) vehicle-to-vehicle (v2v) communications[2], [3] and (ii) vehicle-to-infrastructure (v2i) communica-tions [4]–[7]. In the meantime, the GPS technology hasbeen adopted for navigation purposes at an unprecedentedrate. It is expected that approximately 300 million GPSdevices will be shipped in 2009 alone [8]. It becomes a verytimely topic to develop novel applications by integrating thecutting-edge DSRC and GPS technologies.Specifically, this work is motivated by the observed trend

that a large number of vehicles have started to install GPS-receivers for navigation and the drivers are guided by these

• Jaehoon Jeong, Shuo Guo, Yu Gu, Tian He, and David H.C. Du are with theDepartment of Computer Science and Engineering, University of Minnesota,Twin Cities, 200 Union Street SE, Minneapolis, MN 55455.E-mail: {jjeong,sguo,yugu,tianhe,du}@cs.umn.edu.

GPS-based navigation systems to select better driving pathsin terms of the physically shortest path or the vehicular low-density traffic path. Therefore, the nature research questionis how to make the most of this trend to improve theperformance of vehicular ad hoc networks.Let’s consider the scenario where Internet access points

are sparsely deployed along the roadways for (i) the road-side reports, such as the time-critical reports (e.g., drivingaccident and driving hazard), the road traffic statistics (e.g.,vehicle density and vehicle speed) and (ii) the drivers’video/audio data (e.g., video clip, photo and voice/musicrecording). The Internet access points have limited commu-nication coverage, so the vehicles cannot directly transmittheir packets to the Internet access points. To support sucha scenario, the carry-and-forward techniques are proposedfor use by several opportunistic forwarding schemes [4],[9], [10]. In these schemes, vehicles carry or forward packetsprogressively close to an access point by selecting potentialshortest path based on traffic statistics.Without considering individual vehicles’ trajectories, the

previous forwarding schemes [4], [9], [10] can be inefficient,especially in light-traffic road networks (e.g., rural-area roadnetworks). This is because that the probability to forwardpackets to other vehicles at intersections is low in light-traffic road networks and it would be the case that vehiclescarry packets towards the wrong direction, introducingexcessive long delays. In our study, we let a packet carriervehicle select a best next packet carrier with its neighboringvehicles’ trajectories. Also, in our study, we try to reduce theamount of data packets using the unicasting based on thevehicles’ trajectories. Note that in light-traffic road networks,the volume of data traffic may be not necessarily light. Forexample, in the delivery of the video/audio data for driversor road networks, the data volume can be high. If these high-volume data are delivered using broadcasting or Epidemicrouting [9], the packet traffic will be high. Thus, our goal isto reduce the volume of packet traffic through the vehicle-trajectory-based forwarding.This paper, for the first time, proposes a data forwarding

scheme utilizing the vehicles’ trajectory information forlight-traffic road networks. The first challenge is how to usethe trajectory information in a privacy-preserving manner,while improving the data forwarding performance. To re-

Digital Object Indentifier 10.1109/TPDS.2010.103 1045-9219/10/$26.00 © 2010 IEEE

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMSThis article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

Authorized licensed use limited to: University of Minnesota. Downloaded on July 10,2010 at 04:11:14 UTC from IEEE Xplore. Restrictions apply.

solve this challenge, we design a local algorithm to computeexpected data delivery delay (EDD) at individual vehiclesto an access point, using private trajectory information andknown traffic statistics. Only the computed delay is sharedwith neighboring vehicles. The vehicle with the shortestEDD is selected as the next packet carrier for its neighboringvehicles. The other challenge is how to model an accurateroad link delay, a delay defined as the time taken for a packetto travel through a road segment using carry-and-forward. Toresolve this challenge, we accurately model road link delay,based on traffic density information obtained from the GPS-based navigation system.Our intellectual contributions are as follows:

• An analytical link delay model for packet deliveryalong a road segment that is much more accurate thanthat of the state-of-the-art solution. Besides serving asa critical building block of our TBD design, this linkdelay model is useful for other VANET designs, suchas the E2E delay estimation in VANET routing proto-cols [11] and the data dissemination through network-wide broadcast.

• An expected E2E delivery delay computation based onindividual vehicle trajectory. The E2E delivery delay isestimated using both vehicular traffic statistics and indi-vidual vehicle trajectory. It turns out that this estimationprovides a more accurate delivery delay, so vehiclescan make better decision on the packet forwarding. Ourtrajectory-based delivery scheme opens a new door of apotential research direction based on vehicle trajectoryin vehicular networks.

• A data forwarding with multiple Internet Access Points(APs). Our earlier work TBD [12] utilizes vehicle trajec-tory information along with vehicular traffic statisticsto further improve communication delay and deliveryprobability for vehicle-to-static destination communica-tions. In TBD [12], the data forwarding is designed onlyfor single AP scenarios, not for multiple AP scenarios.In this paper, we take a step further and provide anefficient solution based on vehicle trajectory for the dataforwarding with multiple APs.

The rest of this paper is organized as follows: Section 2 de-scribes the problem formulation. Section 3 describes our linkdelay model. Section 4 explains the design of the trajectory-based forwarding including the computation of the end-to-end delivery delay. Section 5 evaluates our design. InSection 6, we summarize the related work for vehicualrnetworking. We conclude this paper along with future workin Section 7.

2 PROBLEM FORMULATION

Given a road network with an Internet access point, theresearch problem is to minimize the end-to-end deliverydelay of packets to the Internet access point. In this paper,we focus on one-way data delivery which is useful for thetime-critical reports, such as vehicle accidents, road surface

(a) A Light-Traffic Road Network

(b) A Road Network with Unbal-anced Traffic Density

Fig. 1. Packet Delivery Scenarios

monitoring and driving hazards [13]. We leave two-waydelivery as future work.In this paper, we refer (i) Vehicle trajectory as the moving

path from the vehicle’s starting position to its destinationposition in a road network; (ii) Expected Delivery Delay (EDD)as the expected time taken to deliver a packet generated bya vehicle to an Internet access point via the VANET; (iii)Carry delay as a part of the delivery delay introduced whilea packet is carried by a moving vehicle; (iv) Communicationdelay as a part of the delivery delay introduced while apacket is forwarded among vehicles. Our work is based onthe following four assumptions:

• The geographical location information of packet desti-nations, such as Internet access points (APs), is availableto vehicles. A couple of studies have been done toutilize the Internet access points available on the road-sides [6], [7].

• Vehicles participating in VANET have a wireless com-munication device, such as the Dedicated Short RangeCommunications (DSRC) device [1]. Nowadays manyvehicle vendors, such as GM and Toyota, are planningto install DSRC devices at vehicles [14], [15].

• Vehicles are installed with a GPS-based navigation sys-tem and digital road maps. Traffic statistics, such asvehicle arrival rate λ and average vehicle speed v perroad segment, are available via a commercial naviga-tion service, similar to the one currently provided byGarmin Traffic [16].



• Vehicles know their trajectory by themselves. However,vehicles do not release their trajectory to other vehiclesfor privacy concerns.

It should be noted that in the VANET scenarios, the carrydelay is several orders-of-magnitude longer than the commu-nication delay. For example, a vehicle takes 90 seconds totravel along a road segment of 1 mile with a speed of 40MPH, however, it takes only ten of milliseconds 1 to forwarda packet over the same road segment, even after consideringthe retransmission due to wireless link noise or packetcollision. Thus, since the carry delay is the dominating partof the total delivery delay, in the rest of the paper we focuson the carry delay for the sake of clarity, although the smallcommunication delay does exist in our design.Let’s consider the following packet forwarding scenarios

in Figure 1. The first scenario, as shown in Figure 1(a), isthat three vehicles, denoted as Source, Carrier-1 and Carrier-2,are moving in a road network. The Source wants to send itspacket to the access point. Carrier-1 and Carrier-2 are withinSource’s communication range. If trajectories are known, it isclear that Sourcewill decide to forward its packets to Carrier-1, since Carrier-1 moves towards the access point. The firstchallenging problem is how to make such a decision whenprivacy-sensitive trajectories are not shared directly.The second scenario, as shown in Figure 1(b), is that

Carrier-1’s trajectory is on the light road traffic path andCarrier-2’s trajectory is on the heavy road traffic path. Inthis case, Source can select Carrier-2 as next carrier andforward its packet to Carrier-2 since Carrier-2 has a highprobability that it can forward Source’s packets to the accesspoint via a communication path consisting of other vehicles.The second challenging problem is how to combine theroad traffic statistics (e.g., density) information with thevehicle trajectory information for better forwarding decisionmaking.In the next sections, we will deal with the two challenges

raised in this section through the Link delay modeling(in Section 3) and the Trajectory-based forwarding (in Sec-tion 4).

3 THE LINK DELAY MODEL

This section analyzes the link delay for one road segmentwith one-way vehicular traffic given the vehicle arrival rateλ, the vehicle speed v and the communication range R; notethat a constant vehicle speed v is used for the link delayanalysis and that the impact of the variable vehicle speedon the link delay will be shown in comparison of simulationresults at the end of this section; the results indicate that ourlink delay model is a good approximation to the simulationresult. We leave the link delay for a two-way road segmentas future work. Three terms for the link delay model aredefined as follows:Definition 1 (Network Component): Let Network Com-

ponent be a group of vehicles that can communicate with

1. Note that the data rate in DSRC [1] is 6∼27 Mbps and transmissionrange can extend to almost 1,000 meters.

jI

1n2n...1−knkn 0n

fl cl

iI

l

Fig. 2. Forwarding Distance lf and Carry Distance lc

0t0 1t 2t 3t 1−kt ktvRt /0 +

1+kt

fl

2T1T 1−kT0T kT

0t0 1t 2t 3t 1−kt kt 1+kt

fl

Fig. 3. Forwarding Distance (lf ) for Vehicle Arrivals

each other via either one-hop or multi-hop communication,that is, a connected ad-hoc network. Figure 2 shows anetwork component consisting of vehicles n1,..., nk.

Definition 2 (Forwarding Distance): Let Forwarding Dis-tance (denoted as lf ) be the physical distance a packet travelsvia wireless communication within a road segment startingfrom the entrance. Figure 2 shows the forwarding distancelf for the network component.

Definition 3 (Carry Distance): Let Carry Distance (de-noted as lc) be the physical distance a packet is carried bya vehicle within a road segment. Figure 2 shows the carrydistance lc of vehicle n1.

Let v be the vehicle speed. By ignoring the small com-munication delay, the link delay dij along a road with thelength of l is the corresponding carry delay. We have,

dij =lcv

where lc = l − lf . (1)

Therefore, the expected link delay E[dij ] is:

E[dij ] = (l − E[lf ])/v. (2)

In Equation 2, in order to obtain the expected link delayE[dij ], we need to derive the expected forwarding distanceE[lf ] first. Clearly the forwarding distance lf equals thecommunication length of the network component that isnear the entrance as shown in Figure 2. To illustrate our



modeling approach, we use Figure 3(a) to explain how theforwarding distance lf change over time under differenttraffic arrival patterns.

• At time t0, vehicle n0 arrives. Since n0 moves at theconstant speed v, the forwarding distance lf increaseslinearly at the rate of v. During the time interval [t0, t0+R/v], no other vehicle arrives, forcing n0 to move out ofthe communication range of Ii. As a result, lf reducesto zero after t0 + R/v.

• At time t1, vehicle n1 arrives. Similarly, the forwardingdistance lf increases linearly at the rate of v. In thiscase, vehicles n2,..., nk arrive at Ii with the inter-arrivaltime less than R/v, forming a network component of kvehicles.

To formally derive E[lf ], we model the forwarding dis-tance lf as the sum of the inter-vehicle distance of vehicleswithin the component at any time. Figure 3(b) shows thecorresponding vehicle arrival times as in Figure 3(a). Letth be the arrival time of the h-th vehicle. Let Th be theinter-arrival interval of the h-th vehicle and the (h + 1)-thvehicle. Th is assumed to be an exponential random variablewith arrival rate λ. This assumption has been shown validin [17], because the Kolmogorov-Smirnov test can accuratelyapproximate the statistics of vehicle inter-arrival time basedon the empirical data for a real roadway into an exponentialdistribution.As shown in Figure 3(b), when the vehicle nk+1 carrier

arrives at tk+1 with an outgoing packet, the forwardingdistance lf is zero if Tk = tk+1−tk > R/v, otherwise lf is thecommunication length of the network component

∑kh=1 Thv

if Tk = tk+1−tk < R/v. We note the expected number of ve-hicle inter-distances (i.e., vTh) within a network componentis the ratio between P [vTh ≤ R] and P [vTh > R], accordingto detailed derivation in Appendix A. Therefore, we obtainE[lf ] for the road segment (Ii, Ij) as follows:

E[lf ] =E[vTh|vTh ≤ R]×P [vTh ≤ R]

P [vTh > R](3)

From (3), we can see that E[lf ] is the multiplication of (i)the average inter-distance of two adjacent vehicles within thesame component and (ii) the ratio of the probability that theinter-distance is not greater than the communication rangeto the probability that the inter-distance is greater than thecommunication range. As the inter-arrival time decreases,this ratio increases, leading to the longer average forwardingdistance; note that as the inter-arrival time decreases, theaverage inter-distance decreases, but the increasing rateof the ratio is much faster. Therefore, this fits well ourintuition that the shorter inter-arrival time, the shorter inter-distance for communication, leading to the longer averageforwarding distance.Figure 4 shows the average forwarding distance lf com-

parison among simulation model and two analytical modelsfor one-way roadway: (i) Our TBD link model for finite roadlength in Appendix A and (ii) VADD link model proposedby Zhao and Cao [4]. As shown in Figure 4, our link modelgives very accurate average forwarding distance lf estimates

100

200

300

400

500

600

700

800

900

1000

0 2 4 6 8 10 12 14 16 18 20

Avg

. For

war

ding

Dis

tanc

e[m

]

Vehicle Inter-arrival Time[sec]

SimulationTBD

VADD

Fig. 4. Validation and Comparison of Analytical Models

0

10

20

30

40

50

60

70

80

90

100

0 2 4 6 8 10 12 14 16 18 20

Avg

. Lin

k D

elay

[sec

]


SimulationTBD

VADD

(a) Constant Vehicle Speed with μv = 40MPH

0

10

20

30

40

50

60

70

80

90

100

0 2 4 6 8 10 12 14 16 18 20

Avg

. Lin

k D

elay

[sec

]


SimulationTBD

VADD

(b) Vehicle Speed with N(40, 3.5)MPH

0

10

20

30

40

50

60

70

80

90

100

0 2 4 6 8 10 12 14 16 18 20

Avg

. Lin

k D

elay

[sec

]


SimulationTBD

VADD

(c) Vehicle Speed with N(40, 7)MPH

Fig. 5. Link Delay Comparison for Model Validation

under different inter-arrival intervals. The reason VADD isnot accurate is that VADD considers the sum of the lengths



of all connected vehicles, while missing the fact that only thenetwork component starting from the entrance can actuallybe used for data forwarding.The above modeling process assumes the speed v of

vehicles is constant. Clearly it does not hold well in practice,because for four-lane roadways, the vehicle speed deviationis 6.2 MPH (i.e., 9.98 km/h), according to field study con-ducted by Victor Muchuruza [18]. To investigate how robustour link delay model is, we test the accuracy of our modelunder three different settings: (i) a constant vehicle speed of40 MPH, (ii) a normal speed distribution of N(40, 3.5) and(iii) a normal speed distribution of N(40, 7). Our model iscompared with simulation, which approximates the groundtruth, and VADD [4]. Figure 5 illustrates that as the vehiclespeed deviation is within the realistic bound, the link delayof TBD is closer to the simulation result than that of VADD.

4 TBD: E2E DELAY MODEL AND PROTOCOL

In this section, we explain the design of our trajectory-basedforwarding with three steps: First, we will explain how tocompute the Expected Delivery Delay (EDD) consideringboth vehicular traffic statistics and individual vehicle trajectoryin section 4.1. Second, we will describe how vehicles per-form the data forwarding based on EDD in section 4.2. Last,we will explain how to extend the TBD in the road networkswith multiple Internet access points.

4.1 End-to-End Delay Model

In this section, we model the EDD with a stochastic model [4]for a given road network. We define the road network graphfor the EDD computation as follows:

Definition 4 (Road Network Graph): Let a road net-work graph be the directed graph of G = (V, E), whereV = {v1, v2, ..., vn} is a set of intersections in the roadnetwork and E = [eij ] is a matrix of edge eij for vertices vi

and vj such that eij �= eji. Figure 6 shows a road networkgraph.

1node2node

y trajectors'node1

y trajectors'node2

Fig. 6. Road Network Graph for Data ForwardingTo estimate end-to-end delay, we cannot use the tradi-

tional shortest path algorithms, such as Dijkstra’s shortestpath algorithm. This is because when the packet carrier

arrives at an intersection, it is not guaranteed that it canmeet another vehicle moving towards the most preferreddirection. In this case, the packet carrier needs to determinewhether it can forward its packet to another vehicle movingtowards other preferred directions or has to carry it withitself to the next intersection on its trajectory. In order toconsider all of the possible cases in the forwarding at eachintersection, we formulate the data delivery based on thiscarry-and-forward as the stochastic model.

4.1.1 Expected Delivery Delay at Intersection

In this section, we explain how to compute the EDD at anintersection, using a stochastic model. Suppose that a packetat intersection i is delivered towards intersection j. Let dij

be the expected link delay for edge eij in Equation (2). Wenote the expected delay EDD at an intersection depends onthe forwarding direction (i.e., edge). Therefore, we use Dij

denote the EDD at the intersection i when the edge eij isused as the forwarding edge. We formulate Dij recursivelyas follows:

Dij = dij + E[delivery delay at j by forwarding or carry]

= dij +∑

k∈N(j)

PjkDjk

(4)

where N(j) is the set of neighboring intersections of inter-section j. We use this stochastic model to compute the EDDat intersection i because the packet will be delivered withsome probability to one of outgoing edges at intersectionj. This means that when the carrier of this packet arrivesat intersection j, the next carrier on each outgoing edgetowards intersection k will be met with probability Pjk . Wewill explain how to compute the probability Pjk later.

1,2D

3,2D 7,2D

2,1D

Fig. 7. EDD Computation for Edge e1,2

For example, suppose that as shown in Figure 7, a packetcarried by a vehicle arrives at intersection 1 and is senttowards intersection 2. The EDD of D1,2 denotes the end-to-end delivery delay when the carrier sends its packet to theAP via the edge e1,2. First, it will take d1,2 seconds to delivera packet to the intersection 2 via e1,2. Once the packet arrivesat intersection 2, there are three possible cases to deliverthe packet. In other words, the packet can be forwardedto one of three neighboring intersections (i.e., intersection1, 3 or 7) of intersection 2 with some probability. Let D2,1,D2,3 and D2,7 be the EDDs for three edges e2,1, e2,3 and



e2,7, respectively. We can compute D1,2 using the stochasticmodel in (4) as follows:

D1,2 = d1,2 + P2,1D2,1 + P2,3D2,3 + P2,7D2,7.

Let n be the number of directed edges in the road networkgraph G = (V, E), as shown in Figure 6. We have n variablesof Dij for directed edge eij ∈ E(G). Since we have nvariables and n linear equations of (4), we can solve thislinear system using the Gaussian Elimination algorithm.We start to explain how to compute the probability Pjk in

(4). Pij is defined as the average forwarding probability that apacket at intersection i will be delivered to a vehicle movingtowards the neighboring intersection j.

Contact Probability: Contact Probability is defined as thechance that a vehicle can encounter another vehicle at anintersection. Let R be the communication range. Let vi bethe vehicle speed in the intersection area of intersection iwhich is a circle of radius R; that is, vehicles are consideredpassing through intersection i with the same constant speedvi. This vehicle speed can be measured from vehicular trafficby dividing the communication diameter 2R by the averagetravel time per intersection. This indicates that vehicles maypass through another intersection with a different constantspeed, depending on the model specification (e.g., trafficcondition) for each intersection. For simplicity, we regard thevehicle speed vi as constant. Note that this constant vehiclespeed vi is also used for the computation of the E2E deliverydelay as forwarding guidance and that the impact of thevariable vehicle speed on the performance will be shownin Section 5.5; the results show that our E2E delivery delaymodel is a good indicator for the decision-making on thedata forwarding.Let Ti be the intersection passing duration during which

a vehicle is able to communicate with the vehicles aroundthe intersection i. Clearly, Ti is affected by the vehiclespeed, the communication range, the traffic signal pattern,stop signs, and the queueing delay. In practice, averageTi can be obtained through empirical measurements basedon GPS navigation systems [16]. In this study, we use asimplifying model to calculate Ti by assuming the nominalcommunication range is R and a constant speed is vi.Therefore, Ti = 2R/vi. We note our design can use empiricalTi measurements if available.Let CPij be the contact probability that a packet carrier

in the intersection area of i will meet at least one vehiclemoving towards j for during Ti. Suppose that the vehiclearrival at the directed edge eij is Poisson process withvehicle arrival rate λij . Thus, CPij is computed using thePoisson Process probability as follows:

CPij = 1− e−λijTi . (5)

Note that in order to compute an approximation of realcontact probability, all of the vehicles passing through theintersection i are assumed to have the same constant speedvi. Thus, they have the same contact probability CPij foredge eij ; in Section 5, the simulation results indicate that

our E2E delay model based on this contact probability is agood approximation for data forwarding.Forwarding Probability: At an intersection, forwarding isprobabilistic in nature, therefore a packet is forwardedwith best-effort. Let’s define the forwarding probability as thechance that a packet carrier at intersection i can forwarda packet to another vehicle moving towards one of theneighboring intersections jk for k = 1..m. We note thereis a clear distinction between the contact probability andforwarding probability, because a packet will not be forwardedto a contacted vehicle that moves to a wrong direction.To calculate forwarding probability, we need to sort edges

based on the forward priority. For an intersection i with mforwarding edges eijk

(k = 1...m), we can sort them in non-decreasing order, based on their geographically shortest pathlength from intersection i to a packet destination (i.e., AP)via the edge eijk

. This heuristic is based on the observationthat the edge on the geographically shortest path tends toprovide the shortest delivery path; note that the intersectionmodel of VADD [4] uses the angle between the packet des-tination and the edge for the enumeration, but the smallestangle does not always give the shortest path in the roadnetworks of non-grid topology. Therefore, the forwardingprobability P ′

ijkfor each edge eijk

is computed as follows:

P ′ijk

=

{CPij1 for k = 1,

(∏k−1

s=1 (1− CPijs))CPijk

for k = 2..m.(6)

Now we can define the forwarding success probability Pi

for intersection i as the probability that the packet carrierat intersection i can meet any other vehicle and forwardits packet to the other vehicle during its passing durationTi; this forwarding success probability is computed as Pi =∑m

k=1 P ′ijk

. On the other hand, the case of forwarding failurehappens when the packet carrier at intersection i cannotmeet any other vehicle during its passing duration Ti andthe forwarding failure probability Qi is defined as the proba-bility that this forwarding failure happens; this forwardingfailure probability is computed as Qi =

∏ms=1 (1 − CPijs

) =1 −

∑m

k=1 P ′ijk

= 1 − Pi. Thus, it is clear that the sum ofthe forwarding success probability Pi and the forwardingfailure probability Qi is 1.Conditional Forwarding Probability: Clearly, a packetshould not be forwarded to the edge that is worse thanthe edge the carrier moves toward, therefore, we needto compute the conditional forwarding probability that apacket carrier moving on edge eijh

can forward its packetto another vehicle moving on eijk

, that is, Pijk |ijh=

P [packet is forwarded to eijk|carrier moves on eijh

]. Theconditional forwarding probability Pijk|ijh

is computed asfollows:

Pijk|ijh=

⎧⎪⎨⎪⎩

P ′ijk

for k < h,

1−∑h−1

s=1 P ′ijs

for k = h,

0 for k > h.

(7)

Supposing that the packet carrier goes to edge eijh, our

forwarding rule is that only when the packet carrier meets



another vehicle moving on the better edge eijkin terms of

expected E2E delay (i.e., EDD), it forwards its packet to thevehicle; this is case 1 (k < h) in (7). Otherwise, it tries toforward its packet to another vehicle moving on the sameedge eijh

or carries its packet with itself in the case wherethere exists no heading vehicle on the edge eijh

; this is case2 (k = h) in (7). Also, for case 3 (k > h) in (7) where thecarrier’s edge eijh

has shorter EDD than other edges eijk

for k > h, the packet carrier will not forward its packet toanother vehicle moving on the longer-EDD edges, so theconditional forwarding probability for case 3 is 0. Note thatthe probability of case 1 (k < h) and the probability of case2 (k = h) sums up to 1 because the probability to forwardto edges eijk

for k = 1..h − 1 other than the carrier’s edgeeijh

is∑h−1

s=1 P ′ijs

and the probability to forward or carry tothe carrier’s edge is 1−

∑h−1s=1 P ′

ijs.

Average Forwarding Probability: Finally, we can computethe average forwarding probability Pijk

that a packet ar-riving at intersection i will be delivered to the neighbor-ing intersection jk by either forwarding or carry. In orderto compute Pijk

for the packet-delivered intersection jk,we need the branch probability Bijh

that a packet carrierarriving at intersection i will move to intersection jh forjh ∈ N(i). This branch probability Bijh

can be obtainedfrom the vehicular traffic statistics on the edge eijh

; that is,Bijh

is the ratio of the number of vehicles branching fromintersection i to intersection jh to the number of vehiclesarriving at intersection i during the traffic measurementtime (e.g., 1 hour). Therefore, Pijk

is calculated by Law oftotal probability [19] as follows:

Pijk=

∑jh∈N(i)

BijhPijk |ijh

. (8)

From (8), Pijkis the average of the conditional forward-

ing probability Pijk |ijhwith respect to the packet carrier’s

branch probability Bijhto edge eijh

incident to intersectioni that is the current location of the packet carrier.

7,2'P

1,2'P 3,2'P

Fig. 8. Average Forwarding Probability P2,3 at Intersection 2

For example, as shown in Figure 8, suppose that a packetcarrier is placed at intersection 2 in Figure 6 and moves toone of the neighboring intersections with the correspondingbranch probability B2,j for j = {1, 3, 7}; that is, there arethree directions for the packet carrier to take, such as MovingDirection-1, Moving Direction-2 and Moving Direction-3. Wewant to compute the average forwarding probability P2,3

that the packet carrier will deliver its packet onto edge e2,3.We assume that the ascending order of the shortest path

length from intersection 2 towards the AP via the threeedges is e2,7, e2,3 and e2,1. According to this assumption,the contacting order for packet forwarding is the same (i.e.,e2,7, e2,3 and e2,1) and the forwarding probabilities for thesethree edges are P ′

2,7, P ′2,3 and P ′

2,1, respectively. Therefore,the average forwarding probability P2,3 is computed from(8) as follows:

P2,3 = B2,1P2,3|2,1 + B2,3P2,3|2,3 + B2,7P2,3|2,7

= B2,1P′2,3 + B2,3(1 − P ′

2,7).

Note that (a) P2,3|2,1 = P ′2,3 since the shortest path length

for the carrier’s moving edge e2,1 is longer than that forthe forwarding edge e2,3, so the carrier tries to forward itspackets onto e2,3; (b) P2,3|2,3 = 1 − P ′

2,7 since the shortestpath length for the edge e2,7 has the shortest among thethree edges; (c) P2,3|2,7 = 0 since the shortest path lengthfor the carrier’s moving edge e2,7 is shorter than that for theforwarding edge e2,3, so the carrier does not try to forwardits packets onto e2,3.We note this EDD model computes Dij without consid-

ering the trajectory. For example, if two vehicles node1 andnode2 are placed at the same intersection 1 in Figure 6, theirEDDs towards the same packet-delivered edge e1,2 are thesame with each other. Therefore, only with this intersectionEDD model, the individual vehicle’s trajectory does notaffect the computation of EDD, so we cannot determine tochoose which one as the best next carrier. In the next section,we explain how the vehicle trajectory can be added to theEDD computation.

4.1.2 Expected Delivery Delay based on Trajectory

In this section, we explain how to compute the expected E2Edelivery delay (EDD) based on the vehicle trajectory. A trajec-tory is defined as the moving path from a vehicle’s startingposition to its destination position in a road network;

The main idea of trajectory-based forwarding is to divide thedelivery process recursively into two steps: (i) The packet carryprocess at the current vehicle and (ii) the delivery process after thepacket leaves this vehicle. In the case of light traffic, it is possiblethat a vehicle could carry a packet continuously over multipleedges.

Suppose the packet is with the current vehicle. Thisvehicle will travel along a trajectory denoted by a sequenceof intersections: 1 → 2 → · · · → M . Let Cij be thetotal time taken to carry the packet by the vehicle fromthe intersection i to the intersection j along the trajectory(1 ≤ i ≤ j ≤ M ). Formally, Cij =

∑j−1k=i lk,k+1/v. As a

reminder, P ′mn is the forwarding probability in (6) that the

vehicle at intersection m can forward its packets to anothervehicle moving towards the neighboring intersection n. Asa reminder, P c

mn be the carry probability that the vehiclecannot forward its packet at intersection m, and so has tocarry its packets to the adjacent intersection n. Formally,P c

mn = 1−∏

k∈N(m) P ′mk. The expected end-to-end delay D

at the vehicle is computed as follows:



D =

M∑j=1

(P [a packet is carried from intersection 1 to j]

× (C1j + E[delivery delay at intersection j]))

=

M∑j=1

((

j−1∏h=1

P ch,h+1)× (C1j +

∑k∈N(j)

P ′jkDjk))

(9)

In (9), P [a packet is carried from intersection 1 to j] =∏j−1h=1 P c

h,h+1 is the carry probability along thetrajectory from intersection 1 to the intersection j.E[delivery delay at intersection j] =

∑k∈N(j) P ′

jkDjk isthe expected delivery delay after the packet leaves thecurrent vehicle.

6,1D

2,1D

1,2D

7,2D

3,2D

2,3D

4,3D 8,3D

Fig. 9. EDD Computation for Vehicle Trajectory

For example, as shown in Figure 9, let the trajectory be1 → 2 → 3 in the road network in Figure 6. First, thevehicle at intersection 1 can try to forward the packets tothe neighboring intersections 2 and 6. If it cannot forwardthe packets at the intersection 1, it must carry them by thenext intersection 2. When it arrives at intersection 2, it cantry to forward again. If it cannot forward again, it will carrythe packet to the third intersection 3. At the destination, ifthe vehicle cannot forward, it discards the packets. Withthis scenario, the expected delivery delay D is computed asfollows:

D = P ′1,2D1,2 + P ′

1,6D1,6 + P c1,2(C1,2 + P ′

2,1D2,1 + P ′2,3D2,3

+P ′2,7D2,7) + P c

1,2Pc2,3(C1,3 + P ′

3,2D3,2 + P ′3,4D3,4

+P ′3,8D3,8).

So far, we have explained how to compute the EDD basedon the vehicular traffic statistics and individual vehicletrajectory. In the next section, we will explain how vehiclescan use their EDDs in the packet forwarding process.

4.2 Forwarding Protocol Design

In this section, we describe our design of the TBD forwardingprotocol to perform data forwarding among vehicles in orderto deliver data packets to the destination in the givenroad network. Our TBD forwarding rule is as simple as thefollowing:

Within a connected ad-hoc network, packets are forwarded tothe vehicle with a minimum EDD.

2node3node 1node

5node

4node

9node 8node

6node

7node

)()( 12 nodeEDDnodeEDD <

)()( 32 nodeEDDnodeEDD <

(a) Data Forwarding on Road Segment eij . Ve-hicles node1, node2 and node3 construct aconnected network. Since node2’s EDD is lessthan node1’s and node3’s, the packets of node1

and node3 are forwarded to node2.

2node3node 1node

6node

7node

9node 8node

5node

4node

minimum is )( 8nodeEDDnetwork.connectedin the

(b) Data Forwarding around Intersection j. Ninevehicles from node1 to node9 construct a con-nected network. Since node8’s EDD is min-imum in the connected network, node2 for-wards its packets to node8 via node1 andnode9.

Fig. 10. TBD Forwarding Protocol in VANET

In order to efficiently share the vehicles’ EDD valueswithin a connected ad-hoc network, we can use the state-of-the-art vehicle information diffusion schemes in [20], [21];that is, these vehicles’ EDD values can be efficiently spreadto neighboring vehicles through one of these diffusionschemes, and also along with the EDD value, each vehiclecan piggyback its vehicle identifier (ID), location, movingdirection and vehicle speed onto its beacon message. OurTBD work is focused on the data forwarding based on thevehicle trajectory in light-traffic road networks, so the TBDuses the following simple diffusion operation for the dataforwarding, used by many mobile ad-hoc routing protocols,such as AODV and DSR [22]:

• Step 1: Each vehicle periodically updates its EDD,based on its current position on its trajectory, as dis-cussed in Section 4.1.2. Vehicles exchange their beaconmessages (containing the minimum EDD value and the



next-hop vehicle’s ID) with neighboring vehicles withinone-hop communication range. Note that initially, theminimum EDD value and the next-hop vehicle’s IDare set to the vehicle’s own EDD and the vehicle’s ID,respectively.

• Step 2: When each vehicle receives the routing infor-mation (i.e., the minimum EDD value and the next-hopvehicle’s ID) from its neighboring vehicles within one-hop communication range, it compares its EDD withthe announced minimum EDD. If the minimum EDDis less than its own EDD, it updates and announces itsrouting information with the next-hop vehicle towardsthe minimum EDD vehicle in the ad-hoc network.Otherwise, the vehicle announces the minimum EDDand the next-hop vehicle’s ID with its own EDD andvehicle ID, respectively.

• Step 3: When these vehicles construct a connected ad-hoc network, the routing information (i.e., the mini-mum EDD value and the next-hop vehicle’s ID) canbe spread to the ad-hoc network.

Thus, with this diffusion operation, vehicles with packetsforward their packets to the next-hop towards the minimumEDD vehicle.Figure 10 illustrates our TBD forwarding protocol. Fig-

ure 10(a) shows the data forwarding on road segment eij .Suppose that node1 and node3 are within the communicationrange of node2 and they carry their packets. Therefore node1,node2 and node3 form a connected network. Since node2’sEDD is minimum in this connected network, node1 andnode3 forward their packets to node2. Figure 10(b) shows thedata forwarding around intersection j. When node1 arrivesat intersection j, nine vehicles from node1 to node9 constructa connected network. Since node8’s EDD is minimum inthe connected network, the packets of node2 are forwardedto node8 via node1 and node9. Thus, through the diffusionoperation for the data forwarding, vehicles can forwardtheir packets to the best next-hop vehicle in terms of theminimum EDD in the current connected ad-hoc network.

4.3 Forwarding for Multiple Access Points

In a large-scale road network, multiple Internet AccessPoints (APs) are usually required to accommodate the In-ternet access for vehicles. In this case, vehicles need to sendtheir packets towards one of the APs; this is a kind of anycasttowards the APs. We can easily extend our data forwardingframework for this anycast. The intersection EDD Dij fora directed edge eij towards one of APs is the minimumamong the EDDs towards the APs as follows:

Dij = mink∈AP

Dkij (10)

where AP is the set of APs and Dkij is the EDD for access

point APk. For example, Figure 11 shows the road networkgraph with two access points AP1 and AP2. The EDD D1,2

for edge e1,2 is min {D11,2, D

21,2} where D1

1,2 and D21,2 are

computed using (4), respectively.

1node2node

y trajectors'node1

y trajectors'node2

Fig. 11. Road Network Graph with Two APs

Through (10), we can compute the EDD for each directededge. Finally, we can compute the EDD based on the ve-hicle trajectory using (9) along with this intersection EDDconsidering the anycast. Note that we need to update theforwarding probability P ′

jk and the carry probability P ch,h+1

in (9). Since we have the minimum EDD for each directededge, we can compute the forwarding probability by sortingedges in non-decreasing order, based on the EDDs for theedges. After computing the forwarding probability at eachintersection, we can compute the carry probability.

5 PERFORMANCE EVALUATION

In this section, we evaluate the performance of TBD bycomparing it with a state-of-the-art scheme, VADD (usingthe Direction-First-Probe forwarding protocol proposed in[4]) while for the fairness, our link delay model is usedfor both TBD and VADD. The evaluation is based on thefollowing:

• Performance Metrics: We use (i) average delivery delayand (ii) packet delivery ratio as the performance metrics.

• Parameters: In the performance evaluation, we inves-tigate the impacts of (i) Vehicular traffic density (N ), (ii)Vehicle speed (μv), (iii) Vehicle speed deviation (σv), (iv)Packet Time-To-Live (TTL), and (v) Internet access pointdensity (M ).

Note that the link delay model and E2E delay models inboth TBD and VADD are based on constant vehicle speed(s)given to road networks. These two E2E delay models areused to make a forwarding decision-making metric calledEDD. We investigate the effectiveness of these two forward-ing schemes in terms of performance metrics.A road network with 49 intersections is used in the

simulation and one Internet access point is deployed inthe center of the network. Each vehicle’s movement patternis determined by a Hybrid Mobility model of City SectionMobility model [23] and Manhattan Mobility model [24].From the characteristics of City Section Mobility, the vehiclesare randomly placed at one intersection as start positionamong the intersections on the road network and randomlyselect another intersection as end position. The vehicles move



TABLE 1Simulation Configuration

Parameter DescriptionThe number of intersections is 49.

Road network The area of the road map is 8.25km×9km(i.e., 5.1263miles×5.5923miles).

Communication range R = 200 meters (i.e., 656 feet).Number of vehicles The number N of vehicles moving within

(N) the road network. The default of N is 100.The expiration time of a packet. The

Time-To-Live default TTL is∞; that is, there exists no(TTL) packet drop due to TTL expiration.

v ∼ N(μv , σv) where μv = {20, 25, ...,60}Vehicle speed MPH and σv = {1, 2, ..., 10} MPH.

(v) The maximum and minimum speeds areμv + 3σv and μv − 3σv , respectively.The default of (μv , σv) is (40, 5) MPH.Let du,v be the shortest path distance

Vehicle travel from start position u to end position v inpath length the road network. l ∼ N(μl, σl) where

(l) μl = du,v km and σl = 3 km (1.86miles).

according to the roadways from their start position to theirend position. Also, the vehicles wait for a random waitingtime (e.g., uniformly distributed from 0 to 10 seconds) atintersections in order to allow the impact of stop sign ortraffic signal. From the characteristics of Manhattan Mobil-ity, as shown in Table 1, the vehicle travel path length l fromstart position u to end position v is selected from a normaldistribution N(μl, σl) where μl is the shortest path distancebetween these two positions and σl determines a randomdetour distance; this random detour distance reflects thatall of the vehicles do not necessarily take the shortest pathfrom their start position and their end position. Once thevehicle arrives at their end position, it pauses during arandom waiting time and randomly selects another endposition. Thus, this vehicle travel process is repeated duringthe simulation time, based on the hybrid mobility model.The vehicle speed is generated from a normal distribution

of N(μv, σv) [18], [25], as shown in Table 1. In the simulation,the forwarding probability used in the EDD computation iscomputed using Equations from (5) to (8) in Section 4.1.1,based on the average vehicle speeds to generate vehicle speedsfor every two directions per two-way road segment; thatis, these two average speeds per road segment can bemeasured from vehicular traffic by dividing the road segmentlength by the average travel time over the road segment. Forsimplicity, we let all of the road segments have the samespeed distribution of N(μv, σv) in the road network forthe simulation; note that our design can easily extend thissimulation setting to having the variety of vehicle speeddistributions for road segments.During the simulation, following an exponential distri-

bution with a mean of 5 seconds, packets are dynamicallygenerated from 10 vehicles in the road network. The totalnumber of generated packets is 50,000 and the simulationis continued until all of these packets are either deliveredor dropped due to TTL expiration. The system parametersare selected based on a typical DSRC scenario [1]. Unlessotherwise specified, the default values in Table 1 are used.

5.1 Verification of Probability Model

In this subsection, first of all, we verify the E2E delay modeldiscussed in Section 4.1. In our simulation, for each directededge eij , the vehicle arrival rate (λij), and the branchprobability (Bij) are measured every one hour, based onaccumulated vehicular traffic statistics. With these λij andBij , we compute the contact probability CPij , forwardingprobability P ′

ij , average forwarding probability Pij and theE2E delivery delay Dij . Table 2 shows the update of thesevalues in one edge (i.e., e10,11) over the simulation time,that is, from the 2nd hour to the 7th hour. From this table,it can be observed that these probabilities and the E2Edelay estimate are converged over the time. This observationindicates that the forwarding-related probabilities and the E2Edelay estimate based on the past statistics can be used forthe forwarding-related computation in the near future. Thus,with these verified probabilities and estimates, we computethe trajectory-based E2E delay D in (9) as forwarding metric.

TABLE 2Probabilities and Estimates for E2E Delay over Time

Variable 2hour 3hour 4hour 5hour 6hour 7hourλij 0.008 0.008 0.008 0.008 0.009 0.009Bij 0.260 0.273 0.291 0.284 0.300 0.299

CPij 0.159 0.159 0.168 0.169 0.177 0.180P ′

ij 0.159 0.159 0.168 0.169 0.177 0.180Pij 0.427 0.465 0.456 0.464 0.464 0.425Dij 1303 1367 1500 1555 1577 1581

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1000 2000 3000 4000 5000 6000

% o

f D

elay

(C

DF)

Delivery Delay[sec]

TBDVADD

Fig. 12. CDF Comparison for Delivery Delay

5.2 Forwarding Behavior Comparison

We compare the forwarding behaviors of TBD and VADDwith the cumulative distribution function (CDF) of theactual packet delivery delays. From Figure 12, it is very clearthat TBD has smaller packet delivery delay than VADD. Forany given packet deliver delay, TBD always has a largerCDF value than VADD before they both reach 100% CDF.For example, TBD reaches 90% CDF with a delivery delay ofabout 1,500 seconds while the value for VADD is about 1,800seconds. In other words, on average, the packet deliverydelay for TBD is smaller than that for VADD and we willshow this quantitatively in the following subsections.



0

500

1000

1500

2000

2500

20 40 60 80 100 120 140 160 180 200

Avg

. Del

iver

y D

elay

[sec

]

Number of Vehicles[#vehicles]

TBDVADD

(a) Impact of the Number of Vehicles

0

500

1000

1500

2000

2500

20 25 30 35 40 45 50 55 60

Avg

. Del

iver

y D

elay

[sec

]

Vehicle Speed[MPH]

TBDVADD

(b) Impact of Vehicle Speed

0

500

1000

1500

2000

2500

1 2 3 4 5 6 7 8 9 10

Avg

. Del

iver

y D

elay

[sec

]

Vehicle Speed Deviation[MPH]

TBDVADD

(c) Impact of Vehicle Speed Deviation

Fig. 13. Performance Comparison between TBD and VADD for Infinite TTL

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

20 40 60 80 100 120 140 160 180 200

Avg

. Del

iver

y R

atio

Number of Vehicles[#vehicles]

TBDVADD

(a) Impact of the Number of Vehicles

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

20 25 30 35 40 45 50 55 60

Avg

. Del

iver

y R

atio

Vehicle Speed[MPH]

TBDVADD

(b) Impact of Vehicle Speed

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1 2 3 4 5 6 7 8 9 10

Avg

. Del

iver

y R

atio

Vehicle Speed Deviation[MPH]

TBDVADD

(c) Impact of Vehicle Speed Deviation

Fig. 14. Performance Comparison between TBD and VADD for Finite TTL (TTL = 1800 seconds)

5.3 The Impact of Vehicle Number N

The number of vehicles in the road network determinesthe vehicular traffic density in a road network. In thissubsection, we intend to study how effectively TBD canforward packets towards the access point using individualvehicles’ trajectory information. Through our extensive sim-ulations, we observe that under low vehicular traffic density,TBD significantly outperforms VADD in terms of packetdelivery delay. Figure 13(a) shows the packet delivery delaycomparison between TBD and VADD with varying thenumber of vehicles under low vehicular traffic density. Asshown in Figure 13(a), TBD has smaller packet deliverydelay than VADD at all vehicular densities. As expected,one trend is that the delivery delays in both TBD and VADDdecrease as the number of vehicles increases. This is becausethe more vehicles increase the contact probability amongvehicles and have a higer probability that they will pass theInternet access point. The smallest delay reduction is 7.8%at N = 200 while the largest delay reduction is 16.7% atN = 40. From this figure, it can be seen that in the extremelylight-traffic road networks, such as N = 20, the trajectory inTBD has less contribution (i.e., 15% reduction) than in thecases of not-so-light-traffic road networks, such as N = 40(i.e., 16.7% reduction). This is because when the number ofvehicles is so small, the probability that vehicles can meeteach other is relatively low and also the probability that thecarriers will pass the Internet access point is low.Another important trend is that as the number of vehicles

is increasing (e.g., N ≥ 160), the performance gap betweenTBD and VADD is decreasing. For example, for N = 180,TBD reduces the delivery delay of VADD by 8.9%, butreduces the delivery delay by about 8.5% and 7.8% forN = 180 and N = 200, respectively. This is because thehigher vehicular traffic density provides the higher proba-bility that the packets can be forwarded to the vehicle witha smaller EDD on the road segment or at every intersection;that is, since TBD lets a packet carrier forward its packetsto its preceding vehicle on the road segment with a highprobability, it works in the similar way with VADD. Thus,both will have almost the same performance in an extremelyhigh traffic density. As a result, we found that TBD notonly provides significant better data forwarding quality thanVADD in light-traffic road networks that are targeted by thispaper, but also has smaller packet delivery delay even athigh-traffic conditions.

5.4 The Impact of Vehicle Speed μv

In this subsection, we investigate how the change of meanvehicle speed affects the delivery delay. Figure 13(b) showsthe delivery delay under different mean vehicle speeds. Asshown in the Figure 13(b), for both TBD and VADD, thehigher vehicle speed leads to the shorter delivery delay.This is because the high vehicle speed yields high vehiclearrival rate at each road segment, leading to the shorterdelivery delay. However, at all vehicle speeds, the TBD stilloutperforms VADD.



5.5 The Impact of Vehicle Speed Deviation σv

In this subsection, we investigate the impact of vehicle speeddeviation on the performance. We found that under light-traffic road networks, the probability that vehicles meet eachother is low, so the speed deviation has little impact on thedelivery delay. Figure 13(c) illustrates our observation forthe delivery delay according to the vehicle speed deviationwhen the number of vehicles is N = 100. The deliverydelays in both TBD and VADD are almost the same at all ofthe vehicle speed deviations from 1 MPH to 10 MPH.

5.6 The Impact of Packet Time-To-Live TTL

In this subsection, we investigate the impact of the packet’sTime-To-Live (TTL) on the packet delivery ratio, defined asthe ratio between the number of delivered packets to thenumber of packets generated. We set TTL to 30 minutes(i.e., 1800 seconds) in our simulation; that is, if a packet isnot delivered within 30 minutes hour after its generation, itwill be discarded by a packet carrier.Figure 14(a) shows the delivery ratio comparison between

TBD and VADDwith varying number of vehicles in the roadnetwork. As expected, the larger number of vehicles yieldshigher average delivery ratio. The delivery ratio for TBDis increasing roughly linearly with respect to the numberof vehicles. On the other hand, in VADD, the increase ofthe number of vehicles under the light-traffic does notcontribute much to the increase of delivery ratio. Clearly,we can see even at light-traffic condition, TBD has betterdelivery ratio than VADD. Especially, at N = 40, the deliveryratio for TBD is 7.8% higher than that for VADD.We investigate the impact of vehicle speed on the delivery

ratio in Figure 14(b). We can see at all vehicle speeds,TBD has larger delivery ratio than VADD. However, theperformance difference between two schemes is gettingsmaller as the vehicle speed increases. This is because withhigher vehicle speed, the vehicle arrival rate also increases ateach road segment and this gives VADD a higher forwardingprobability.We also investigate the impact of vehicle speed devia-

tion on the delivery ratio. Figure 14(c) shows the deliveryratio comparison between TBD and VADD according tothe vehicle speed deviation from 1 MPH to 10 MPH. Inthe simulation, the other parameters use the default valuesspecified in Table 1; that is, the number of vehicles is 100and the average vehicle speed is 40 MPH. TBD is overallbetter than VADD with 4.1% more delivery ratio. Also, asdiscussed in Section 5.5, the vehicle speed deviation doesnot affect the delivery ratios of both TBD and VADD.

5.7 The Impact of Internet Access Point Number M

In this subsection, we explain how multiple Internet AccessPoints (APs) have an impact on the performance. Note thatmultiple APs are uniformly placed in the road network inthe simulation. The other parameters are set to the defaultvalues in Table 1; that is, the number of vehicles is N = 20;note that this is an extremely light-traffic density.

0

500

1000

1500

2000

2500

1 2 3 4 5 6 7 8 9

Avg

. Del

iver

y D

elay

[sec

]

Number of APs[#APs]

TBDVADD

(a) #APs vs. Delivery Delay

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1 2 3 4 5 6 7 8 9

Avg

. Del

iver

y R

atio

Number of APs[#APs]

TBDVADD

(b) #APs vs. Delivery Ratio

Fig. 15. Performance Comparison between TBD and VADDaccording to the Number of APs

Figure 15 shows the performance comparison betweenTBD and VADD according to the number of APs in termsof delivery delay and delivery ratio. As observed in Fig-ure 15(a), the delivery delays for both TBD and VADD aredramatically decreasing and the delivery ratio are quicklyincreasing as the number of APs increases. TBD outperformsVADD by reducing the VADD’s delivery delay by morethan 10% up to 5 APs. On the other hand, as shown inFigure 15(b), for the delivery ratio, TBD has at least 5%better ratio than VADD under a low AP density (e.g., upto 2 APs). However, both have the similar ratio (i.e., morethan 98%) under a high AP density (e.g., from 6 to 10 APs).Through the performance evaluation, we can conclude

that by using the vehicle trajectory information, TBD canprovide better data delivery than VADD in light-trafficvehicular networks at a variety of settings in terms of thevehicular traffic density, vehicle speed distribution, and APdensity.

6 RELATED WORK

Data forwarding and data access issues in VANET havegained a lot of attention recently [2], [4], [5], [26]–[29].The data forwarding in VANET is different from that inthe traditional mobile ad-hoc networks (MANETs) [22]for the reason of (i) vehicles are moving on the physicallyconstrained areas (i.e., roadways), (ii) the moving speed isalso limited by the speed limit on the roadways and (iii)the communication shortest path does not always matchthe physical shortest path due to heterogeneous vehiculartraffic conditions on road segments. These unique charac-teristics of the road networks open the doors of research



opportunities for the data forwarding in the VANET. Also,the frequent network partition and mergence due to the highmobility make the MANET routing protocols [22] ineffectivein the VANET settings [17]. Thus, in order to deal withthis frequent network partition and mergence, the carry-and-forward approaches are necessary in the Disruption TolerantNetworks (DTN), such as VANET.Epidemic Routing in [9] is an early work to support the

data forwarding in the DTN consisting of mobile nodes, de-signed for two-dimensional open fields, not for the vehicularnetworks with the confined routes for vehicles. It allowsthe random pair-wise exchange of data packets amongmobile nodes in order to maximize the possibility that datapackets can be delivered to their destination node. Thus,multiple copies of data packets exist during the delivery.This scheme is effective to deliver the data packets forcritical messages (e.g., accidents) to the destination (e.g., AP)as soon as possible. On the other hand, our TBD investigateshow effectively to deliver data packets with both vehicles’trajectories and vehicular traffic statistics; our TBD is basedon the unicasting, that is, with one copy of a data packeton-the-fly rather than multiple copies.Data forwarding schemes investigating the layout of road

network and vehicular traffic statistics are proposed inVADD [4] and Delay-Bounded Routing [5]. VADD investi-gates the data forwarding using a stochastic model basedon vehicular traffic statistics in order to achieve the lowestdelivery delay from a mobile vehicle to a stationary packetdestination. On the other hand, Delay-Bounded Routingproposes data forwarding schemes to satisfy the user-defineddelay bound rather than the lowest delivery delay. In addition,it also aims at minimizing the channel utilization in terms ofthe number of packet transmissions. Our TBD, in contrast,improves forwarding performance by utilizing the vehicletrajectory information along with vehicular traffic statistics inorder to compute the accurate expected delivery delay forbetter forwarding decision making.MDDV [26] proposes a forwarding scheme in VANET to

allow the predefined packet trajectory. The packet trajectoryin this scheme is the path where this packet traversesthrough, and so is different from the vehicle trajectory.Since this scheme forces the packet to traverse through thepredefined path, it can be inefficient in the light-traffic roadnetworks. This is because the probability that no vehiclemoves along a road segment that is on the edge of packettrajectory is high in the light-traffic road networks.For dense road networks, such as urban roadways, CAR,

MMR and VVR are proposed [2], [27], [28]. CAR forwardsdata packets through the connected path from the packetsource to the packet destination. In rural roadways whichis our focus in this paper, this connectivity-based dataforwarding may not work well due to the sparse vehiculartraffic. MMR and VVR use greedy forwarding choosing thenext packet carrier based on the geographical proximitytowards the packet destination. However, in road networks,since the vehicular traffic distribution is not uniform, thisgeographical greedy forwarding does not always provide

the communication shortest path. On the other hand, ourTBD allows a packet carrier to choose the best next packetcarrier on the communication shortest path since it is awareof the road-network-wide vehicular traffic density alongwith individual vehicle trajectory.In [7], Bychkovsk et al. show the possibility that vehicles

can access open WiFi access points for the Internet access invehicular networks. Cabernet [6] proposes one-hop Internetaccess schemes using open WiFi access points in vehicularnetwork, whose target is different from TBD’s, that is, themulti-hop ad-hoc networking in VANET.OPERA [30] investigates an opportunistic packet replay-

ing using two-way vehicular traffic on road segments. Thedata packets forwarded in one direction on a road segmentcan advance towards the endpoint of the road segmentusing other vehicles moving in the opposite direction. Onthe other hand, our TBD uses only vehicles moving in thesame direction on the road segment for the clarity of thelink delay modeling. The link delay model based on two-way vehicular traffic is left as future work.

7 CONCLUSION

In this paper, we propose a trajectory-based data forwardingscheme for light-traffic road networks, where the carrydelay is the dominating factor for the end-to-end deliverydelay. We compute the aggregated end-to-end carry delayusing the individual vehicle trajectory along with the vehic-ular traffic statistics. Our design allows vehicles to sharetheir trajectory information without exposing their actualtrajectory to neighbor vehicles. This privacy-preserving tra-jectory sharing scheme is made possible by exchangingonly the expected delay value using local vehicle trajectoryinformation. We also propose a link delay model based onthe common assumption of exponential vehicle inter-arrivaltime. It is shown to be more accurate than the state-of-the-art solution.With the increasing popularity of vehicular ad-hoc net-

working, we believe that our forwarding scheme opens afirst door for exploiting the potential benefit of the vehicletrajectory for the performance of VANET networking. Asfuture work, we will develop a data forwarding schemefrom stationary nodes (i.e., Internet access points) to movingvehicles for supporting the Infrastructure-to-Vehicle datadelivery in vehicular networks. This reverse forwarding tomoving vehicles is needed to deliver the road conditioninformation such as the bumps and holes for the drivingsafety. However, this reverse data forwarding is a morechallenging problem because we need to consider boththe destination vehicle’s mobility and the packet deliverydelay. Also, we will investigate the impact of data trafficvolume on the trajectory-based data forwarding in light-traffic vehicular networks and develop a data forwardingscheme considering the data traffic volume, the vehicletrajectory, and the vehicle contact time for communicationsalong the road segment.



ACKNOWLEDGMENT

This research is supported in part by NSF grants CNS-0626609, CNS-0626614, and CNS-0720465. We also receivethe facility support from MSI and DTC at University ofMinnesota.

REFERENCES

[1] A. Carter, “The Status of Vehicle-to-Vehicle Communication as aMeans of Improving Crash Prevention Performance,” Tech. Rep. 05-0264, 2005, http://www-nrd.nhtsa.dot.gov/pdf/nrd-01/esv/esv19/05-0264-W.pdf.

[2] V. Naumov and T. R. Gross, “Connectivity-Aware Routing (CAR) inVehicular Ad Hoc Networks,” in INFOCOM. IEEE, May 2007.

[3] Q. Xu, R. Sengupta, and D. Jiang, “Design and Analysis of HighwaySafety Communication Protocol in 5.9 GHz Dedicated Short RangeCommunication Spectrum,” in VTC. IEEE, Apr. 2003.

[4] J. Zhao and G. Cao, “VADD: Vehicle-Assisted Data Delivery in Ve-hicular Ad Hoc Networks,” IEEE Transactions on Vehicular Technology,vol. 57, no. 3, pp. 1910–1922, May 2008.

[5] A. Skordylis and N. Trigoni, “Delay-bounded Routing in VehicularAd-hoc Networks,” in MobiHoc. ACM, May 2008.

[6] J. Eriksson, H. Balakrishnan, and S. Madden, “Cabernet: VehicularContent Delivery Using WiFi,” in MOBICOM. ACM, Sep. 2008.

[7] V. Bychkovsky, B. Hull, A. Miu, H. Balakrishnan, and S. Madden, “AMeasurement Study of Vehicular Internet Access Using In Situ Wi-FiNetworks,” in MOBICOM. ACM, Sep. 2006.

[8] H. Yomogita, “Mobile GPS Accelerates Chip Development,” http://techon.nikkeibp.co.jp/article/HONSHI/20070424/131605/.

[9] A. Vahdat and D. Becker, “Epidemic Routing for Partially-connectedAd Hoc Networks,” Tech. Rep., 2000, http://www.cs.duke.edu/∼vahdat/ps/epidemic.pdf.

[10] L. Pelusi, A. Passarella, and M. Conti, “Opportunistic Networking:Data Forwarding in Disconnected Mobile Ad Hoc Networks,” IEEECommunications Magazine, vol. 44, no. 11, pp. 134–141, Nov. 2006.

[11] Y. Ding, C. Wang, and L. Xiao, “A Static-Node Assisted AdaptiveRouting Protocol in Vehicular Networks,” in VANET. ACM, Sep.2007.

[12] J. Jeong, S. Guo, Y. Gu, T. He, and D. Du, “TBD: Trajectory-Based DataForwarding for Light-Traffic Vehicular Networks,” in ICDCS. IEEE,Jun. 2009.

[13] B. Hull, V. Bychkovsky, Y. Zhang, K. Chen, M. Goraczko, A. Miu,E. Shih, H. Balakrishnan, and S. Madden, “Cartel: A DistributedMobile Sensor Computing System,” in SenSys. ACM, Nov. 2006.

[14] General Motors (GM), “Vehicle-to-Vehicle (V2V) Communications,”http://www.gm.com/experience/technology/research/overview/isl/vcim.jsp.

[15] Toyota Motor Corporation (TMC), “TMC Develops Onboard DSRCUnit to Improve Traffic Safety,” http://www2.toyota.co.jp/en/news/09/09/0903.html.

[16] Garmin Ltd., “Garmin Traffic,” http://www8.garmin.com/traffic/.[17] N. Wisitpongphan, F. Bai, P. Mudalige, and O. K. Tonguz, “On the

Routing Problem in Disconnected Vehicular Ad Hoc Networks,” inINFOCOM Mini-symposia. IEEE, May 2007.

[18] V. Muchuruza and R. Mussa, “Traffic Operation and Safety Analysesof Minimum Speed Limits on Florida Rural Interstate Highways,”in Proceedings of the 2005 Mid-Continent Transportation Research Sympo-sium, Ames, Iowa, USA, Aug. 2005.

[19] M. DeGroot and M. Schervish, Probability and Statistics (3rd Edition).Addison-Wesley, 2001.

[20] T. Nadeem, S. Dashtinezhad, C. Liao, and L. Iftode, “TrafficView:Traffic Data Dissemination Using Car-to-Car Communication,” MobileComputing and Communications Review (MC2R), Special Issue on MobileData Management, vol. 8, no. 3, pp. 6–19, Jul. 2004.

[21] L. Wischhof, A. Ebner, and H. Rohling, “Information Disseminationin Self-Organizing Intervehicle Networks,” Transactions on IntelligentTransportation Systems, vol. 6, no. 1, pp. 90–101, Mar. 2005.

[22] E. M. Royer and C.-K. Toh, “A Review of Current Routing Protocolsfor Ad-hoc Mobile Wireless Networks,” IEEE Personal Communications,vol. 6, no. 2, pp. 46–55, 1999.

[23] T. Camp, J. Boleng, and V. Davies, “A Survey of Mobility Modelsfor Ad Hoc Network Research,” Wireless Communications and MobilityComputing (WCMC): Special Issue on Mobile Ad Hoc Networking: Re-search, Trends and Applications, vol. 2, pp. 483–502, 2002.

[24] F. Bai, N. Sadagopan, and A. Helmy, “IMPORTANT: A frameworkto systematically analyze the Impact of Mobility on Performance ofRouTing protocols for Adhoc NeTworks,” in INFOCOM. IEEE, Mar.2003.

[25] D. S. Berry and D. M. Belmont, “Distribution of Vehicle Speeds andTravel Times,” in Proceedings of the Second Berkeley Symposium onMathematical Statistics and Probability, 1951.

[26] H. Wu, R. Fujimoto, R. Guensler, and M. Hunter, “MDDV: A Mobility-Centric Data Dissemination Algorithm for Vehicular Networks,” inVANET. ACM, Oct. 2004.

[27] Y. Lee, H. Lee, N. Choi, Y. Choi, and T. Kwon, “Macro-level andMicro-level Routing (MMR) for Urban Vehicular Ad Hoc Networks,”in GLOBECOM. IEEE, Nov. 2007.

[28] H. Lee, Y. Lee, T. Kwon, and Y. Choi, “Virtual Vertex Routing (VVR)for Course-Based Vehicular Ad Hoc Networks,” in WCNC. IEEE,Mar. 2007.

[29] P. Rodriguez, R. Chakravorty, J. Chesterfield, I. Pratt, and S. Banerjee,“MAR: A Commuter Router Infrastructure for the Mobile Internet,”in Mobisys’04, 2004.

[30] M. Abuelela, S. Olariu, and I. Stojmenovic, “OPERA: OpportunisticPacket Relaying in Disconnected Vehicular Ad Hoc Networks,” inMASS. IEEE, Sep. 2008.

APPENDIX ALINK DELAY MODELING

A.1 Average Forwarding Distance for Infinite Road Seg-ment

The E[lf ] can be computed as the expected sum of the inter-distances within the network component. Suppose that theinter-arrival time Th is exponentially distributed with arrivalrate λ. So Th for h = 1..k are i.i.d. for the exponentialdistribution with parameter λ. Let a = R/v; that is, a is thetime taken for a vehicle to move out of the communicationrange R with speed v. Let C(k) be the condition for thecomponent consisting of k vehicle inter-arrivals such thatC(k): T0 > a and Th ≤ a for h = 1..k. Let L(k) be the lengthof the network component of k vehicle inter-arrivals. Then,E[lf ] can be derived using the law of total expectation asfollows:

E[lf ] = E[L] =∞∑

k=1

E[L(k)|C(k)]× P [C(k)]

= v ×

∞∑k=1

E[

k∑h=1

Th|T0 > a, Th ≤ a for h = 1..k]

× P [T0 > a, Th ≤ a for h = 1..k]

(11)

Since, in (11), Th for h = 0..k are i.i.d. for the exponentialdistribution with parameter λ, we can rewrite (11) as fol-lows:

E[lf ] = v×

∞∑k=1

k×E[Th|Th ≤ a]×P [Th ≤ a]k×P [T0 > a] (12)

Since P [Th ≤ a] = 1 − e−λa and P [T0 > a] = e−λa,respectively, we need to compute E[Th|Th ≤ a] to compute



(12):

E[Th|Th ≤ a] =

∫ a

0

t× P [Th = t|Th ≤ a]dt

=

∫ a

0

t×P [Th = t, Th ≤ a]

P [Th ≤ a]dt

=

∫ a

0

t×P [Th = t]

P [Th ≤ a]dt

=

∫ a

0

t×λe−λt

1− e−λadt

=1/λ− (a + 1/λ)e−λa

1− e−λa.

(13)

Therefore, (12) can be rewritten as follows:

E[lf ] = α×

∞∑k=1

kβk−1 (14)

where α = ve−λa( 1λ− (a + 1

λ)e−λa) and β = 1− e−λa.

Let f(β) =∑∞

k=1 βk. Since 0 < β < 1, f(β) = β1−β

.Accordingly, f ′(β) = d

dk(∑∞

k=1 βk) =∑∞

k=1 kβk−1 = 1(1−β)2 .

Therefore, E[lf ] is as follows:

E[lf ] =α

(1− β)2= v

1/λ− (a + 1/λ)e−λa

1− e−λa×

1− e−λa

e−λa(15)

Since P [Th ≤ a] = 1− e−λa and P [T0 > a] = e−λa, we have

E[lf ] = vE[Th|Th ≤ a]×P [Th ≤ a]

P [Th > a]

= E[vTh|vTh ≤ R]×P [vTh ≤ R]

P [vTh > R].

(16)

A.2 Average Forwarding Distance for Finite Road Seg-ment

For the finite road length l, we need to guarantee that thecomponent length must be less than or equal to the roadsegment length. The idea is to let the component lengthL′(k) ≤ l using function min along with L(k) for the infiniteroad length as follows:

L′(k) = min{l, L(k)} where L(k) = v

k∑h=1

Th. (17)

(11) can be rewritten using (17) as follows:

E[lf ] =∞∑

k=1

E[L′(k)|C(k)]× P [C(k)]

=

∞∑k=1

E[L′(k)|T0 > a, Th ≤ a for h = 1..k]

× P [T0 > a, Th ≤ a for h = 1..k]

=

N−1∑k=1

E[L(k)|T0 > a, Th ≤ a for h = 1..k]

× P [T0 > a, Th ≤ a for h = 1..k]

+

∞∑k=N

l × P [T0 > a, Th ≤ a for h = 1..k]

(18)

In (18), we need to determine N which is the index to letthe component length longer than the road length l. Letg(k) = E[L(k)|C(k)]. We can compute g(k) as follows:

g(k) = vk × E[Th|Th ≤ a]

= vk ×1/λ− (a + 1/λ)e−λa

1− e−λa

=α

β(1− β)× k

where α = ve−λa(1

λ− (a +

1

λ)e−λa) and β = 1− e−λa.

(19)

We can search the smallest positive integer N to satisfyg(N) ≥ l with (19) as follows:

α

β(1 − β)×N ≥ l ⇒ N = �

β(1 − β)

αl. (20)

In the similar way with (14), we can compute the summa-tions of (18) using the differential of f(β) =

∑N−1k=1 βk =

β−βN

1−β. Therefore, E[lf ] is as follows:

E[lf ] =α((N − 1)βN −NβN−1 + 1)

(1− β)2+ lβN (21)

where α = ve−λa( 1λ− (a + 1

λ)e−λa) and β = 1 − e−λa, and

N = �β(1−β)α

l.

Jaehoon (Paul) Jeong is currently a software en-gineer in Brocade Communications Systems. Hereceived the Ph.D. degree under Professor DavidH.C. Du and Professor Tian He from the Depart-ment of Computer Science and Engineering at theUniversity of Minnesota in 2009. He received theB.S. degree from the Department of Information En-gineering at Sungkyunkwan University in Korea andthe M.S. degree under Professor Yanghee Choi fromthe School of Computer Science and Engineeringat Seoul National University in Korea, in 1999 and

2001, respectively. Also, he was a researcher in Electronics and Telecommu-nications Research Institute (ETRI) in Korea from 2001 to 2004. In ETRI, heresearched on the address auto-configuration and DNS systems for mobilead-hoc networks, and also participated in the Internet Standardization in theInternet Engineering Task Force (IETF), such as IPv6 DNS Configuration.He has published two IETF standards of RFC 4339 and RFC 5006 for IPv6DNS Configuration. His current research interests are the wireless sensornetworking for road networks and the data forwarding in vehicular networks.Dr. Jeong is a member of IEEE, ACM and IETF.

Shuo Guo received her B.S. in Electronic Engineer-ing at Tsinghua University in 2006 and is currentlya Ph.D. candidate in the Department of Electricaland Computer Engineering at the University of Min-nesota, Twin Cities. Her research includes WirelessSensor Networks, Vehicular Ad-Hoc Networks, andReal-time and Embedded Systems. She receiveda best paper award at IEEE MASS 2008 and haspublication in many premier sensor network journalsand conferences.



Yu (Jason) Gu is currently a Ph.D. candidate in theDepartment of Computer Science and Engineeringat the University of Minnesota. He is the author andco-author of over 21 papers in premier journals andconferences. His publications have been selected asgraduate-level course materials by over 10 univer-sities in the United States and other countries. Hehas received several prestigious awards from theUniversity of Minnesota, including a Doctoral Disser-tation Fellowship from 2008 to 2009, an Excellencein Research Award in 2007, a General Dynamics

Research Fellowship in 2006 and an Academic Excellence Fellowship Awardin 2006. His research includes Networked Embedded Systems, WirelessSensor Networks, Cyber-Physical Systems, Wireless Networking, Real-timeand Embedded Systems, Distributed Systems, Vehicular Ad-Hoc Networksand Stream Computing Systems. Yu Gu is a member of ACM, IEEE andSIAM.

Tian He is currently an assistant professor in theDepartment of Computer Science and Engineeringat the University of Minnesota-Twin City. He receivedthe Ph.D. degree under Professor John A. Stankovicfrom the University of Virginia, Virginia in 2004. Dr.He is the author and co-author of over 90 papersin premier sensor network journals and conferenceswith over 4000 citations. His publications have beenselected as graduate-level course materials by over50 universities in the United States and other coun-tries. Dr. He has received a number of research

awards in the area of sensor networking, including four best paper awards(MSN 2006 and SASN 2006, MASS 2008, MDM 2009). Dr. He is alsothe recipient of the NSF CAREER Award 2009 and McKnight Land-GrantProfessorship 2009-2011. Dr. He served a few program chair positionsin international conferences and on many program committees, and alsocurrently serves as an editorial board member for four international journalsincluding ACM Transactions on Sensor Networks. His research includeswireless sensor networks, intelligent transportation systems, real-time em-bedded systems and distributed systems, supported by National ScienceFoundation and other agencies. Dr. He is a member of ACM and IEEE.

David H.C. Du received the B.S. degree in mathe-matics from National Tsing-Hua University, Taiwan,R.O.C. in 1974, and the M.S. and Ph.D. degrees incomputer science from the University of Washing-ton, Seattle, in 1980 and 1981, respectively. He iscurrently the Qwest Chair Professor at the ComputerScience and Engineering Department, University ofMinnesota, Minneapolis. His research interests in-clude cyber security, sensor networks, multimediacomputing, storage systems, high-speed networking,high-performance computing over clusters of work-

stations, database design and CAD for VLSI circuits. He has authored andco-authored more than 210 technical papers, including 100 referred journalpublications in his research areas. He has also graduated 49 Ph.D. and80 M.S. students. Dr. Du is an IEEE Fellow and a Fellow of MinnesotaSupercomputer Institute. He is currently serving a number of journal editorialboards. He has also served as guest editors for a number of journalsincluding IEEE Computer, IEEE and Communications of ACM. He hasalso served as Conference Chair and Program Committee Chair to severalconferences in multimedia, database and networking areas. Most recently,he is the General Chair for IEEE Security and Privacy Symposium (Oakland,California) 2009 and Program Committee Co-Chair for International Confer-ence on Parallel Processing 2009.



Documents

Trajectory-Based Data Forwarding for Light-Trafﬁc … Data Forwarding for Light-Trafﬁc Vehicular Ad-Hoc Networks Jaehoon Jeong, Student Member, IEEE, Shuo Guo, Student Member,